Designing for the middle means losing your best and your brightest [burnsteem100]
Ever since @primevaldad mentioned it to me a week or two ago, I've been fascinated by NotebookLM. This week, I used it to make an AI narrated podcast to describe my recent post, Addressing Steem's fundamental social reward challenge: One size fits all - fits nobody.
The podcast "video" is here: The Fighter Jet Fix to Steem's Rewards.
Unfortunately, it's 17 minutes long, so I couldn't upload to YouTube, and that means no embedding here. I think it's worth listening to, though, so I encourage you to click through - and you can listen on double speed to get through it in less than 10. (The photo in the video is meaningless, I just grabbed a hawk picture that I had handy, because it had to include something.)
It's impressive to see how the AI was able to understand concepts that are mostly unique to Steem, and it was also able to weave in relevant facts from the rest of the Internet, and even imagery from real-world human experience.
To be fair, the AI also got a number of facts wrong, engaged in a lot of speculation, assumed facts not in evidence, and it overstated many arguments. However, I think it got most of the important points from my article right.
It really impresses me that the AI came up with this quote at the end of the video:
Look at your work place, your community, or the systems you participate in every day. What rigid, average, reward structures are currently in place, and who are the top-tier talents being quietly driven away by those structures without you even realizing it? ... designing for the middle means losing your best and your brightest.
Anyway, click through to listen to the video - The Fighter Jet Fix to Steem's Rewards,
And you can read the original article that it's summarizing, here: Addressing Steem's fundamental social reward challenge: One size fits all - fits nobody
Update (May 9, 2026): After multiple attempts failed at generating a YouTube sized video, it finally succeeded. Here's a shorter, ~8 minute video, where embedding works:
Thank you for your attention!
All photos above were taken by me with a Nikon P1000 camera. I'm sharing them under the CC BY 4.0 license (share & adapt freely with attribution to the original source).
Know of any blockchain audio/video sites?
I think I hosted on of my deep dive episodes on 3Speak recently.
I haven't used it, but there is Speak on Steem (speem.watch). If I understand right, they have options to host the videos on YouTube or else on IPFS.
Not sure what 3Speak uses nowadays, since they were hostile to Steem back at the time of the Hive fork, and I lost track of them after that. I was also going to mention dlive, but now I see that they shut down last month. There's also, maybe, d.tube. For a while, it looked like their site was dead, but now it has videos again. Not sure what to make of that.
Oh, I knew of d.tube. But I also thought it was dead.
What did you use to put your audio and image together? Is that just a feature Rumble offers?
I've been trying to find something open source / free where I can put a slide show and generated audio together, and I'm not happy with what I've tried so far.
I might also try the video generation with NotebookLM; my past attempts haven't been as good as the audio, but it's been worked on a lot since the last time I tried. I'm sure it's better now.
I had to hunt around for a while. I wound up using Movie Maker, but I don't edit video enough to know if it's any good. It doesn't feel very intuitive...
I was happy with the results, if/when it finishes successfully, but it fails a alot.
This seems a little odd because a selling point early on for NLM was that it only (read: largely—it still needs some common sense context that it gets from training data) pulls from sources you provide.
That doesn't altogether prevent it from hallucinating. I've seen some doozies in my experimentation! But it should stick relatively to the context at hand via sources you've uploaded.
I wonder if they've relaxed that a bit to make a more viable product overall.
They extended the number of sources you can add from 50 to 100, but even with that, I often wish I could add more sources just for context to point it in the right direction.
Great to see the results of your experimentation!
Generating a "Deep Dive" audio episode is a new favorite way to learn for me.
I haven't started using it, but they've integrated NLM into Gemini now so you can reference your notebook in your chat interface. "Generate a cover image for the article about how focusing on middles leads to losing the best."
I can definitely understand that, although based on my other comment, "assumes facts not in evidence", it could steer you wrong if you don't compare against the source documents. Maybe that tendency is reduced as you add additional sources, though.
I think it's also helpful to shape a general understanding if you acknowledge that details might be off. To my understanding, general orientation is an excellent use case for LLMs. They are mathematically locked in to a source-specific semantic space.
So, if you're looking for a general overview and understanding to get oriented in a space, it works well. If you're looking for research grade details to support a particular thesis, look elsewhere! (For now; classical programming can be used for complementary tools to guide the LLM in the background, and I think that's starting to become a standard environment set-up).
Great to see the results of your experimentation!
Generating a "Deep Dive" audio episode is a new favorite way to learn for me.
I haven't started using it, but they've integrated NLM into Gemini now so you can reference your notebook in your chat interface. "Generate a cover image for the article about how focusing on middles leads to losing the best."